What the documents look like
Site selection for a chain headquarters happens before a store opens. The core work is lease negotiation and construction delivery, not in-store supervision. The working materials are a site evaluation form, a lease letter of intent, a renovation budget and an opening schedule. Their data arrives from scattered sources: trade-area footfall, map POI data, POS history by store type, and landlord documents. Updates are irregular and occur once per store.
Operating reviews and exception reviews must attach each figure and each conclusion back to a specific document version. The rent-to-sales ratio, the delivery milestones and the deposit amount lose their verifiable anchor if a review cites a number without the source text and its status. A lease letter of intent is not a signed lease. A rent quotation with a validity date is not a locked contract term. The review has to keep these stages separate.
Constraints at the citation tracing step
The knowledge base search node selects one or more knowledge bases that use the same vector model, for vector search. Its output is an array of cited content whose length may be zero. Even when nothing is found, that output path still runs. An empty citation array therefore does not mean the workflow has errored or stopped, and the downstream step must be designed to accept an empty array.
The minimum relevance value is a filter threshold, effective only in semantic search or when result reranking is used. It is not a final sorting rule. After question optimization, hybrid search, image retrieval or result reranking are enabled, the final result may come from multiple recall paths fused together and does not necessarily follow a single vector similarity order. RRF is the fusion step for hybrid search results, not a switch that can be turned on. Reranking requires a ReRank model and must be enabled in settings. Image retrieval requires corresponding image input and model capability. These optional paths do not all execute on every request. Knowledge bases with different vector models cannot be associated together; the model identifier must be checked first and non-matching bases excluded. A cross-model search must not be run to compare missed recalls.
Hypothetical paper exercise: a chain headquarters expansion department prepares, on a hypothetical date of 10 March 2025, a site evaluation form. The source text reads: "Proposed leased property: Unit 12, Jia Road; owner: a property company; intended lease term: five years; first-year rent quotation: RMB 800,000; quotation valid until 31 March 2025." On the same date the lease letter of intent reads: "The parties have reached a preliminary intention regarding Unit 12, Jia Road; specific terms are subject to the formal lease contract; this contract has not been signed." A draft review sentence to be assessed reads: "The lease contract for Unit 12, Jia Road has taken effect and the first-year rent of RMB 800,000 is locked." Authorization boundary: these materials are used only for internal review checking and do not indicate that any signing, approval or payment action has been completed.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Corresponding action for the sample |
|---|---|---|
| Associated knowledge bases | Select one or more knowledge bases with the same vector model | Check that the vector model identifiers of the bases holding the site evaluation form and the lease letter of intent match |
| Search mode | Semantic search, full-text search or hybrid search | For a proper name such as "Unit 12, Jia Road", full-text or hybrid search may be selected |
| Citation limit | Each search cites at most n tokens of content | For this hypothetical trial the starting value is set at 2000 tokens; the reason is that the sample material is short, and the value is compared with the actual sample length before adjustment |
| Minimum relevance | A 0-1 value, effective only in semantic search or result reranking | For this hypothetical trial the starting value is set at 0.3, compared with the sample recall results to judge whether too much is filtered |
| Result reranking | Requires a ReRank model, enabled in settings | If enabled, check whether the rerank score corresponds to the cited source text |
| Citation output | Array format, length may be zero | Use a paper input of [] to check the downstream design |
Three common errors
- Treating a retrieval hit as authorization evidence. In the sample, the lease letter of intent states "this contract has not been signed". Retrieving that sentence does not mean the contract has taken effect. The comparison object is the signing-status wording in the source text; the observable result is whether the draft sentence matches that status. Here it does not, so the draft sentence is unsupported and must be marked for verification.
- Treating minimum relevance as a permission or entity filter. That value only filters low-relevance results; it does not exclude unrelated businesses or records without access rights. The input is the same query with different relevance values; the comparison is whether the recalled content still contains records outside the access boundary; the observable result is that adjusting relevance does not change the access boundary.
- Treating an empty citation output as a workflow failure. The node output length may be zero and the path still runs. The input is an explicit paper empty array; the comparison is whether the downstream node continues with an empty array; the observable result is that the downstream step does not stop because of an empty citation.
How to verify the setup
- Check each cited source against the source text. The input is the sample source text together with the draft sentence "The lease contract for Unit 12, Jia Road has taken effect and the first-year rent of RMB 800,000 is locked." Compare the source wording "this contract has not been signed" with the contract-status wording of the draft sentence. The observable result is that the two statuses do not match, so the draft sentence is unsupported and must be marked for verification.
- Check the quotation validity field. The input is the source wording "quotation valid until 31 March 2025." Compare whether the draft sentence presents that date as a supply period or a contract term. The observable result is that a quotation validity date does not indicate a supply period, and conclusions beyond that field's scope are unsupported.
- Check empty-citation handling. The input is a paper empty array []. Compare whether the downstream node design declares that it can handle a citation length of zero. The observable result is that the downstream step continues with an empty array and does not treat the empty citation as an error termination. All of the above are paper expectations and were not actually run.